Fix importing unofficial TF models with extra optimizer weights

[monologg]

Hi:)

I was trying to convert the BERT tf model to torch model, and tf model has extra optimizer weights (This file is the one I've tried to convert).

But it encounters the error, and I've printed the parameters' name in tf model.

I've found out the instead of "adam_v" for "adam_m, names are saved as "AdamWeightDecayOptimizer" or "AdamWeightDecayOptimizer_1".

There was a similar issue that also encountered the issue that I had. (Issue Link from DeepPavlov repo)

I can't find out the exact reason why the parameters are named as "AdamWeightDecayOptimizer" or "AdamWeightDecayOptimizer_1" instead of "adam_v" or "adam_m", but it might be safe to cover all the exceptional cases:)

[LysandreJik]

Looks good to me

[codecov-io]

Codecov Report

Merging #2652 into master will not change coverage.
The diff coverage is 0%.

@@           Coverage Diff           @@
##           master    #2652   +/-   ##
=======================================
  Coverage   74.58%   74.58%           
=======================================
  Files          87       87           
  Lines       14892    14892           
=======================================
  Hits        11107    11107           
  Misses       3785     3785

Impacted Files	Coverage Δ
src/transformers/modeling_bert.py	87.9% <0%> (ø)	⬆️
src/transformers/modeling_albert.py	79.14% <0%> (ø)	⬆️
src/transformers/modeling_t5.py	81.09% <0%> (ø)	⬆️

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update babd41e...d338eb0. Read the comment docs.